Have the default target default to not using the Julia runtime.

[maleadt]

Otherwise the kernel example fails on macOS running Julia 1.9:

Reason: unsupported call through a literal pointer (call to pthread_getspecific)

This is because the TLS getter is lowered differently:

define void @_Z17julia_kernel_2628() #0 !dbg !4 {
top:
  %0 = call {}*** @julia.get_pgcstack()
  %1 = bitcast {}*** %0 to {}**
  %current_task = getelementptr inbounds {}*, {}** %1, i64 -13
  %2 = bitcast {}** %current_task to i64*
  %world_age = getelementptr inbounds i64, i64* %2, i64 14
  %3 = bitcast {}*** %0 to {}**
  %current_task1 = getelementptr inbounds {}*, {}** %3, i64 -13
  %ptls_field = getelementptr inbounds {}*, {}** %current_task1, i64 15
  %ptls_load = load {}*, {}** %ptls_field, align 8, !tbaa !7
  %ptls = bitcast {}* %ptls_load to {}**
  fence syncscope("singlethread") seq_cst
  %4 = bitcast {}** %ptls to i64**
  %5 = getelementptr inbounds i64*, i64** %4, i64 2
  %safepoint = load i64*, i64** %5, align 8, !tbaa !11, !invariant.load !6
  %6 = load volatile i64, i64* %safepoint, align 8
  fence syncscope("singlethread") seq_cst
  ret void, !dbg !13
}

optimizes to

define void @_Z17julia_kernel_2997() local_unnamed_addr #0 !dbg !5 {
top:
  %0 = call {}*** inttoptr (i64 7114247836 to {}*** (i64)*)(i64 261) #1
  %ptls_field3 = getelementptr inbounds {}**, {}*** %0, i64 2
  %1 = bitcast {}*** %ptls_field3 to i64***
  %ptls_load45 = load i64**, i64*** %1, align 8, !tbaa !8
  fence syncscope("singlethread") seq_cst
  %2 = getelementptr inbounds i64*, i64** %ptls_load45, i64 2
  %safepoint = load i64*, i64** %2, align 8, !tbaa !12
  %3 = load volatile i64, i64* %safepoint, align 8
  fence syncscope("singlethread") seq_cst
  ret void, !dbg !14
}

while on other platforms:

define void @_Z17julia_kernel_2198() #0 !dbg !4 {
top:
  %0 = call {}*** @julia.get_pgcstack()
  %1 = bitcast {}*** %0 to {}**
  %current_task = getelementptr inbounds {}*, {}** %1, i64 -13
  %2 = bitcast {}** %current_task to i64*
  %world_age = getelementptr inbounds i64, i64* %2, i64 14
  %3 = bitcast {}*** %0 to {}**
  %current_task1 = getelementptr inbounds {}*, {}** %3, i64 -13
  %ptls_field = getelementptr inbounds {}*, {}** %current_task1, i64 15
  %ptls_load = load {}*, {}** %ptls_field, align 8, !tbaa !7
  %ptls = bitcast {}* %ptls_load to {}**
  fence syncscope("singlethread") seq_cst
  %4 = bitcast {}** %ptls to i64**
  %5 = getelementptr inbounds i64*, i64** %4, i64 2
  %safepoint = load i64*, i64** %5, align 8, !tbaa !11, !invariant.load !6
  %6 = load volatile i64, i64* %safepoint, align 8
  fence syncscope("singlethread") seq_cst
  ret void, !dbg !13
}

... optimizes to

define void @_Z17julia_kernel_2936() local_unnamed_addr #0 !dbg !5 {
top:
  %thread_ptr = call i8* asm "movq %fs:0, $0", "=r"() #1
  %ppgcstack_i8 = getelementptr i8, i8* %thread_ptr, i64 -8
  %ppgcstack = bitcast i8* %ppgcstack_i8 to {}****
  %pgcstack = load {}***, {}**** %ppgcstack, align 8
  %ptls_field3 = getelementptr inbounds {}**, {}*** %pgcstack, i64 2
  %0 = bitcast {}*** %ptls_field3 to i64***
  %ptls_load45 = load i64**, i64*** %0, align 8, !tbaa !8
  fence syncscope("singlethread") seq_cst
  %1 = getelementptr inbounds i64*, i64** %ptls_load45, i64 2
  %safepoint = load i64*, i64** %1, align 8, !tbaa !12
  %2 = load volatile i64, i64* %safepoint, align 8
  fence syncscope("singlethread") seq_cst
  ret void, !dbg !14
}

which doesn't trigger the validator (inline asm vs call-to-pointer)

---

on 1.8, where we didn't emit a safepoint at the start, we used to get:

define void @_Z17julia_kernel_2728() #0 !dbg !5 {
top:
  %0 = call {}*** @julia.get_pgcstack()
  %1 = bitcast {}*** %0 to {}**
  %current_task = getelementptr inbounds {}*, {}** %1, i64 -12
  %2 = bitcast {}** %current_task to i64*
  %world_age = getelementptr inbounds i64, i64* %2, i64 13
  ret void, !dbg !7
}